講演情報
[4O1-IS-2a-03]Multi-Modal Loan Default Prediction via LLM-Based Text–Table Consistency Modelling
〇Haoming Zhang1, Tengfei Shao1, Masayuki Goto1 (1. Waseda University)
regular
キーワード：
Multimodal Consistency、Large Language Model、Risk Assessment、P2P Lending
In peer-to-peer (P2P) lending, effective risk assessment requires integrating quantitative financial records with qualitative borrower narratives, as the latter convey critical intent signals that structured data alone cannot capture. However, existing multimodal approaches typically fuse these modalities as complementary inputs, largely overlooking the semantic discrepancies between a borrower’s subjective narrative and their objective financial reality. We argue that such misalignments serve as potent behavioral proxies for default risk, revealing hidden signals of overconfidence or strategic misrepresentation. To address this, we propose a discrepancy-aware framework leveraging Large Language Models (LLMs) to perform granular consistency reasoning. Unlike traditional embedding-based fusion, our method explicitly quantifies feature-level contradictions. The evaluation experiments are conducted by using the data of Lending Club to show the effectiveness of the proposed approach. The experiments demonstrate that integrating these discrepancy signals significantly outperforms existing baselines in default prediction, while offering superior interpretability regarding the source of risk.